Practical notes from the engineers who build agents, copilots and AI products.
Self-hosted AI agents pay back through three lines: hours removed from a repeatable process, per-token spend that stops leaving your organisation, and outcomes only possible once data stays inside the perimeter. A defensible case counts the first, discounts the second and treats the third separately.
Retrieval-augmented generation fixes missing knowledge; fine-tuning fixes missing behaviour. Most products need retrieval first, fine-tuning rarely, and both only when an evaluation proves a specific failure survives prompt and retrieval improvements.
Plan twelve to eighteen weeks from kick-off to an agent acting on its own: an eight to sixteen week build plus a shadow period. A single-task agent on cloud GPUs can be live in eight weeks. Owned hardware, several write integrations or a regulated review cycle push you to the top of the range.
Self-hosted AI agent projects fail for five reasons: the model is chosen before the evaluation set exists, capacity is sized for the average not the peak, tools carry broad write access, the agent goes live without shadow running, and nobody owns it after launch. Each is a decision, not an accident.
A self-hosted AI agents RFP should name the workflows, the systems each agent must reach, where the model runs and on whose hardware, the acceptance thresholds you will test against, the ownership and exit terms, and a costed first year. Anything less produces quotes you cannot compare.
The hidden costs of self-hosted AI agents are idle GPU capacity, integration work against systems nobody owns, the security review, building and rerunning evaluation suites, model refresh when open weights are superseded, and monthly care. None of the six is optional and most quotes price only the build.
For a startup, self-hosted AI agents are usually one workflow on one box, chosen for unit economics or a customer's data clause. For an enterprise, they are a governed platform touching many systems under audit. The technology is the same; the scope, the approval path and the cost are not.
In India a self-hosted AI agent programme runs from ₹20,80,000 to ₹72,00,000 plus infrastructure, billed in rupees with GST, and takes eight to sixteen weeks. Delivery model, residency and who operates the GPUs matter more to the final number than the model you choose.
Ask how they will prove the agent works, who owns the code, prompts and weights, what the first twelve months cost including infrastructure, what happens when an agent acts wrongly, and how you leave. Vendors who have shipped self-hosted agents answer all five without preparation.